1. Bureau of Transportation Statistics
Alternative Approaches for Measuring the Movement of Goods in the United States
Julie Parker
Bureau of Transportation Statistics
October 27, 2018

2. Commodity Flow Survey (CFS)
A little background. Also, Auxiliary establishments are included such as warehouses and distribution centers.
Commodity Flow Survey (CFS)
Component of the Economic Census
Every 5 years
Partnership: Bureau Transportation Statistics (BTS) and US Census Bureau
Shipper survey of ~100,000 establishments
Mining
Manufacturing
Wholesale Trade
Select Retail and Services

3. CFS Data and Usage Information collected: Commodity Value Weight
Mode of Transportation
Destination
Temperature Control, HAZMAT, Exports
Data provide a comprehensive, multimodal picture of national freight flows for sampled industries.

4. CFS Process for Respondent
Sample shipments
Questionnaire
Extract shipment data
US Census Bureau
Industry Respondent
Census receives data
Enter data into online questionnaire
Find commodity codes
They do this 4 times since reporting is quarterly for the 1 year.

5. CFS, Past & Present Started in 1993 Many good changes since 1993:
Website
Optical character recognition
Electronic Questionnaire
Methodology
Publish millions of rows of data

6. Technological Advances
1993
2017

7. Computational power
AND,

8. Room for Improvement
1993
2017
Room for improvement especially with data collection.

9. Reduce Respondent Burden
Sample shipments
Questionnaire
Extract shipment data
US Census Bureau
Industry Respondent
Census receives data
Enter data into online questionnaire
Find commodity codes
I’m going to talk about how these 2 items will reduce respondent burden.

10. Reduce Respondent Burden
Sample shipments
Questionnaire
Extract shipment data
US Census Bureau
Industry Respondent
Census receives data
Enter data into online questionnaire
Find commodity codes
I’ll first talk about searching for commodities.

11. Standard Classification of Transported Goods (SCTG)
As you can see from the questionnaire, we ask for Standard Classification of Transported Goods (SCTG) and a description. Respondents essentially take that description and match it to 1 of 514 SCTG code. What also makes this more painful is that these codes are not used by industry. But, we need some kind of classification method to produce estimates.
Standard Classification of Transported Goods (SCTG)
514 SCTG codes
Not used by Industries
Classification needed

12. Standard Classification of Transported Goods (SCTG)
1 sec - Use a segue to next slide.
Standard Classification of Transported Goods (SCTG)
514 SCTG codes
Not used by Industries
Classification needed

13. SCTG Resources Census Business Help Site – Search Tool
Here are some SCTG resources – PDF and online search tool (was made ad hoc). Use the pig/hog example.
SCTG Resources
Census Business Help Site – Search Tool

14. Objective: Text Classification
Create a Machine Learning Model
Assign SCTG code to shipment records
Semi-supervised – because uses some unlabeled data (doesn’t have SCTG); classifier model

15. SCTG Machine Learning Model
Consider all variables that benefit the algorithm
SCTG Machine Learning Model
Training Data: 360,000 from 500,000 shipments
6.4 million shipments from 2017 CFS
Machine Learning / Classification Algorithms: Natural Language Processing (Bag-of-Words), Logistic Regression (logit) Model, Grid Search CV
Features: 31 Variables, 514 SCTG Labels
Current model: SCTG code, Shipment Description, Industry code

16. SCTG ML Model Results Model Performance: 76%
“Truth” is respondent’s reported SCTG
Model is doing a better job!
Easing Respondent Burden in 2022 CFS:
2022 CFS, respondents save 50,000+ hours, $2.1 million in respondent lookup costs
Ongoing: more cleaning, relabeling, scraping new product descriptions, creating better process

17. Industry Results Accuracy varies by industry MANUFACTURING
We need to looks at this from an industry-level since that’s how we sample. As you can see, accuracy varies by industry. For example, manufacturing can match to 70% of the descriptions for that industry. Similar to the model performance, these estimates are effected by respondents’ reported SCTG being the truth deck.

18. CFS analyst team uses it as an aid, to look up potential SCTG codes for a given NAICS/description combination
Added bonus, 2017 Edit tool
Look up potential SCTG codes for a given industry/description combination

19. Reduce Respondent Burden
Sample shipments
Questionnaire
Extract shipment data
US Census Bureau
Industry Respondent
Census receives data
Enter data into online questionnaire
Find commodity codes
Line separates Census/Respondent. It’s very difficult for businesses to find these commodities codes.

20. Provide Data from Internal Systems
Targeting Large Companies/Enterprises
3 Approaches:
BTS Research
Company-Led
CFS Team Initiated

21. Approach 1: BTS Research
What’s Industry doing?
Outreach, attend their conferences
How’s data managed?
Are they providing data to someone else?
Supply Chain Intelligence

22. Establishment Software System
Enterprise resource planning (ERP) 
Business process management software 
Integrates all facets of an operation single database, application and user interface
How are these data different from what we currently get?
It’s an integrated management system (talk about value and mode) used by large establishments – especially enterprises. There are many vendors – Oracle, SAP, Microsoft, etc.

23. Transforming Freight Flow Data Collection
This example is 1 of 3 projects that was awarded.
Transforming Freight Flow Data Collection
Using business-to-business (B2B) data to improve CFS use as time series enabling more frequent updates and interim period reporting
ERP system
Recruiting case studies
Grocery store chain
Paper company

24. Approach 2: Company-Led
CURRENTLY: Reporting is one location at a time
Separate user names and passwords
RECENTLY: Companies are looking to reduce their burden.
Data dumps
Set up internal system
‘Click a button’
Companies know there’s a better way. Click a button and the companies data (for multi-establishments) are sent to the government.

25. Approach 3: CFS Team Initiated
Site visit to understand how company pulls data
Nonresponse follow-up
Understand reporting issues
2022 CFS, ‘set up their system’

26. Questions and Comments
(202)
Special thanks to:
Joy Sharp
James Hinckley
Mehdi Hashemipour
Christian Moscardi

27. Lastly,
Will respondents provide more data, more often, if burden is reduced? Other incentives needed? Mandatory reporting is very helpful but how far will it take us?